Add and maintain AI experts
Start with curated responsibilities and skill combinations to build a team of specialist partners. Reuse experts across spaces while choosing who handles the host, SDK, testing or guides and which skills they use.
Outcome: Inspect a role, prepare selected skills, confirm the effective configuration and add the expert to a space.
Before you begin: Use a desktop source build with the collaboration workbench. Prepare the model, experts and practice projects described below; check release notes for installer coverage.
On this page
01Inspect a role before adding it
Open Collaboration workbench → My experts → Add expert. Choose Choose from role library or Design your own role. Role details show responsibilities, methods, deliverables, boundaries and skills. Browsing or closing details does not install skills or create an effective expert. Choose to add only when ready.
02Step 1: Define role
A template prefills the role. Designing your own opens the three-step configuration directly; an AI conversation is not required. Enter a name, specialization, responsibilities, deliverables, methods and boundaries, with an optional preferred model. Use assessable work such as “Maintain the host–SDK interface contract, including validation and error responses,” rather than personality adjectives alone.
03Step 2: Configure skills
Official templates initially select their reviewed default combination. Keep or adjust it. Add or replace skills offers Role recommendations, All skills and source import. Inspect purpose, provenance, fixed version and role-specific reasons before selecting. Selection means the expert uses the skill; it is independent of installation status. Removing a declaration does not uninstall the shared global package.
| Example role | Methods in its default combination |
|---|---|
| Product manager | create-prd, user-stories, prioritize-features: requirements, stories and priorities. |
| Architect | architecture-patterns, api-design-principles, architecture-decision-records: architecture, interfaces and decisions. |
| QA engineer | test-scenarios, webapp-testing, verification-before-completion: scenarios, browser checks and delivery verification. |
04Step 3: Confirm and prepare
Review the name, responsibilities, skills and usage notes. Filled template steps are directly accessible; you do not need to repeatedly choose Next. Complete required fields, then choose Add expert and prepare skills. The host reuses matching fixed packages and prepares missing items with actual progress. The configuration takes effect after all items succeed. Installation proves availability for loading, not acceptance of a business result.
05Recover from preparation failure
Leave the progress page and return through the expert card. Inspect the failed item, resolve network, catalog expiry, unavailable release or disabled-skill issues, then retry or adjust the selection. You decide whether a disabled global skill should be enabled. Do not substitute an unverified source merely because its name matches. A failed edit keeps the previous effective configuration.
06Edit configurations and drafts
Open Edit expert profile from the profile or card menu. Existing experts use a single page for basic information, responsibilities and skills. Unchanged content needs no save. When leaving actual edits, choose whether to save a draft, discard or continue. Added skills may need preparation; removing declarations keeps shared packages installed. Refreshing the role library does not restore defaults over your choices. Started tasks retain their expert and skill snapshots.
07Add the expert to a space
Assign the prepared expert to a space and describe its responsibility for this goal. Roles and skills grant no additional file, tool or project permissions. Historical sources remain limited to authorized scope; reusing an expert does not mean it has read every project. Native subagents are temporary delegates within a task and do not automatically become reusable library experts.
Link experts, projects and a goal08Catalog and version scope
Curated roles and fixed skills have recorded real installation tests against the test Store. Current availability is determined by the catalog verified by the host. This does not imply open community publishing or large-scale automated review. If new artifacts cannot be obtained, keep effective configurations and retry when the source is available. A role appearing on this page does not mean it is installed.
Check the result and troubleshoot
Inspect a role, prepare selected skills, confirm the effective configuration and add the expert to a space.
Check actual results using this page’s verification steps. For business integrations, inspect UI submission, tool receipts and server records separately.
Check states and recovery steps →